Version 0.1.2, draft for review. Wire version token: gidp/0.1.
Most of what an economy would like to match never gets stated. A company that would sell at the right price does not announce it. A candidate open to the right role does not tell their employer. A fund that would follow a credible lead does not publish the conditions. The interest is real, the conditions are knowable, and stating them publicly costs more than the match is worth.
Graduated Interest Disclosure is an open protocol for that class of interest. It lets an agent, acting for a principal under an explicit and auditable delegation, establish whether two parties are worth introducing, without either side disclosing what it is protecting and without a central operator holding both sides’ secrets.
The primitive is the Conditional Interest: not necessarily seeking X, but authorising the Agent to explore X if conditions C1…Cn hold, without revealing protected information until disclosure conditions D1…Dn are satisfied. The protocol defines how such an interest is expressed, what may be exposed and to whom, what two agents may ask each other, what they must answer truthfully, and where the exchange stops and a human takes over.
GIDP is a discovery layer. It does not negotiate, does not settle, does not custody anything, and does not commit a principal to anything. It hands a qualified, consented Opportunity to a human or to a downstream protocol, and stops there.
| Path | What it is |
|---|---|
spec/GIDP-0.1.md |
The specification. Normative. |
spec/explainer.md |
One page, for a first reader. |
spec/use-cases.md |
Eight cases the protocol is meant to serve. |
spec/rationale.md |
Why the rules are what they are: histories, worked failures, borrowed distinctions. Non-normative. |
spec/open-problems.md |
The formal open problem, the questions 0.1 knowingly defers, and the positions taken. Non-normative. |
spec/alternatives.md |
Why not use a broker, a listing, a secure comparison or a set intersection? Measured, not argued. Non-normative. |
spec/LIMITS.md |
Where this protocol should not be used. Non-normative. |
spec/CHANGELOG.md |
Versioning policy and history. |
impl/ |
Reference implementation in Python, with the four worked domains, the probing attack, the four limit cases, and the JSON Schemas. |
impl/SPEC-ISSUES.md |
Every ambiguity the implementation found in the draft, and what changed because of it. |
Read the explainer first, then the specification. If you are looking for
reasons to reject the design, spec/LIMITS.md, spec/alternatives.md and
spec/open-problems.md are where the weaknesses are written down,
including the measurement showing that one of the specification’s own
stated mitigations does not work, and the number saying how much a probing
counterparty can extract.
cd impl
python -m pip install -e .
python -m pytest # the conformance and scenario suite
python examples/cross_border.py
python examples/probing.py # what an adversarial querent can extract
python examples/limits.py # the four cases chosen because they should break
python -m baselines.compare # the same case through four existing mechanisms
python -m baselines.mitigations # do the protocol's own defences work?
python -m baselines.sybil # what a query budget is worth against many identities
python -m baselines.projection # the Section 11.4 trade-off, measured
The implementation tests the specification and has no authority over it: where
the two disagree, the specification is wrong until it is fixed, and the
disagreement is logged in impl/SPEC-ISSUES.md.
0.1 is a draft published for review, not a standard. The most useful review is
an attempt to break it: a market it claims to serve and does not, an inference
attack cheaper than the one in examples/probing.py, a reading of the
truthfulness rule that permits a lie, or a second implementation that disagrees
with this one. Known implementations, including a first independent
clean-room implementation and what it found, are listed in
IMPLEMENTATIONS.md.
The probing attack has been measured. Unprotected, a counterparty that spends questions instead of asking once
extracts 65% of the worked case’s hypothesis space over its five private
facts (9.94 of 15.30 bits, as spec/alternatives.md defines them), in seventeen claims,
and the coarsening the specification illustrates stops none of it. Two
controls the specification recommends do bound it (§24.3). An information
budget kept per Standing Interest, and decided on the worst case over the
answers it might give, caps what any probing counterparty can extract at the
budget (two bits of a five-bit secret in the measured case; the bisection
measured there obtained under one) while serving most honest counterparties; because it is kept
per interest rather than per counterparty, an adversary holding several
identities does not dilute it. A claim lattice caps the resolution of any
answer, whatever the number of claims. A budget kept per
counterparty, by contrast, is bypassed by multiple identities (§24.4).
spec/alternatives.md has the numbers.
What remains open is narrower. A budget can be spent by
an adversary, so that honest counterparties are refused: a cost in
availability rather than in secrecy. Correlations between attributes are not
bounded by budgets kept attribute by attribute. And no result yet says how
good the best truthful policy can be. That is the open problem of §24.3: bounding
what an adaptive querent learns from a truthful compatibility oracle while
keeping it useful. Candidate directions are in
spec/open-problems.md, including one that the limit cases have since
qualified.
Documents under spec/, and this README: CC BY 4.0; see
LICENSE-DOCS.md.
Code under impl/: Apache License 2.0; see LICENSE.
Florian Williams, Independent. contact@gidp.dev. Canonical location: https://gidp.dev. Source and issue tracker: https://github.com/FlorianWilliams/gidp.